Agencies Eye Bulk Token Buying as AI Compute Costs Reshape Media Models
The Development
Digiday's latest Bold Call, published March 2, 2026, argues that AI compute costs are becoming the defining cost-structure question for agencies—and the next battleground in upfront negotiations and principal media arrangements. As AI handles more agency workflows—copywriting, performance reporting, agentic task execution—token consumption is emerging as a new, highly variable cost center that neither agencies nor clients currently know how to forecast. The model being floated: agencies buy tokens in bulk directly from AI companies, absorb the variable usage costs, and package the capability into client subscriptions at a fixed rate. Digiday reporters Sam Bradley and Seb Joseph noted that marketers are already quietly inserting contract clauses demanding transparency around compute costs, signaling that the tension around this model is surfacing faster than agencies anticipated.
Our Take
The parallel to TV upfronts is structurally accurate but undersells the transparency problem. When agencies buy TV inventory as principals, the margin debate is uncomfortable but bounded—inventory has a known market price. Token costs have no such anchor. An agency absorbing compute risk and reselling it as a subscription has near-total opacity on markup, and clients have almost no benchmark to push back against. This is exactly the environment where principal media abuses thrived before scrutiny forced disclosure. The contract clause trend is the tell: sophisticated clients already sense the information asymmetry. Agencies moving fastest to lock in token bulk deals are building a margin structure that will face the same advertiser backlash that reshaped programmatic principal media—just on a faster timeline because AI spend is scaling inside existing client relationships, not through new budget lines.
The Exploit
Action Item
If you're a marketing procurement lead renewing or negotiating an agency contract in Q1 or Q2 2026, demand an explicit AI compute cost passthrough clause that requires itemized token usage reporting before signing—before bulk-deal structures become standard boilerplate that's far harder to unwind.